Deep Cogito v2 ships 70B/109B/405B/671B open-weight family with Iterated Distillation & Amplification self-improvement loop
Deep Cogito's v2 release ships four open-weight sizes (70B, 109B, 405B, 671B) wired into an Iterated Distillation & Amplification (IDA) self-improvement loop. The release positions IDA as a deployable architecture rather than a research curiosity — the first open-weight family where the "model improves itself between checkpoints" methodology is shipped as the default training recipe.
IDA is the long-standing alignment-research proposal where a model amplifies its own reasoning under supervision and then distills the amplified reasoning back into the base weights. Until Cogito v2, the methodology lived primarily in Anthropic and DeepMind research papers. Cogito v2 is the first attempt to ship it as a production training pipeline for an open-weight family — which means the methodology is now publicly inspectable, criticizable, and replicable.
For the responsible-scaling debate, the implication is double-edged. Self-improvement loops are exactly the class of architecture the AISI red-team frameworks were designed to monitor. Cogito releasing the methodology open means regulators get the inspection surface they wanted — and also means the methodology gets adopted by labs without responsible-scaling commitments.